In preview

What Databricks is shipping but has not settled. Kept on its own page on purpose: a preview is a thing to investigate, not a thing to build a pipeline or an exam answer on.

51 tracked · 12 written up
PreviewPublic Preview22

Open to everyone and supported for production use, though it can still change. On by default, and an admin can turn it off.

BetaBeta25

Open to most customers but not for production: support runs through engineering, and the shape of it can still move. On by default on Premium, off on Enterprise.

PrivatePrivate Preview4

Invite only, through your account team, and not for production. Here so you know it exists, not so you can plan around it.

Preview Radar

51 of 51 tracked
Public PreviewBetaPrivate PreviewGAFoundationsWorkspaceSQLData engineeringAI / MLThe ai_search SQL functionConfigurable recovery peri…Domains and subdomainsGit CLI commands in Git fo…Lakehouse ReplayServerless usage policiesABAC DENY policiesEnvironments on classic co…Fine-grained DML privilegesGenie Code task in jobsGovernance HubPagesREPLACE USING flowsSharing foreign schemas an…Unit testing for pipelinesBundle deployment historyPrivate network gatewayServerless compute rate li…Dashboard relationshipsGenie InspectGenie OntologyQuery tagsRelationship-scoped and cr…Lakehouse Real-Time wareho…Memory in Genie OneWeb search in GenieAnomaly detectionAutomatic type widening in…Managed Iceberg materializ…MySQL and PostgreSQL conne…Real-time mode in Lakeflow…Zerobus Ingest into defaul…Lakeflow Jobs health scoreAI RuntimeAI Runtime and serverless…Databricks-managed MCP ser…Feature viewsMLflow deployment jobsMLflow deployment jobsAgent and MCP services as…Agent services in Unity Ca…Full-text search indexes i…Managed agent memoryMLflow Prompt RegistryMLflow review queuesProduction monitoring for…Production monitoring for…Prompt registryService policies for AI se…Unified trace tableUnity Catalog SkillsThe ai_search SQL function — SQL, Beta1Configurable recovery period for dropped tables — Catalog, Public Preview2Domains and subdomains — Catalog, Public Preview3Git CLI commands in Git folders — Workspace, Public Preview4Lakehouse Replay — Compute, Public Preview5Serverless usage policies — Workspace, Public Preview6ABAC DENY policies — Catalog, Beta7Environments on classic compute — Compute, Beta8Fine-grained DML privileges — Catalog, Beta9Genie Code task in jobs — Jobs & Pipelines, Beta10Governance Hub — Workspace, Beta11Pages — Catalog, Beta12REPLACE USING flows — Jobs & Pipelines, Beta13Sharing foreign schemas and tables — Discover, Beta14Unit testing for pipelines — Jobs & Pipelines, Beta15Bundle deployment history — Workspace, Private Preview16Private network gateway — Workspace, Private Preview17Serverless compute rate limits — Compute, Private Preview18Dashboard relationships — Dashboards, Public Preview19Genie Inspect — Genie Agents, Public Preview20Genie Ontology — Genie Agents, Public Preview21Query tags — Query History, Public Preview22Relationship-scoped and cross-dataset filters — Dashboards, Public Preview23Lakehouse Real-Time warehouse — SQL Warehouses, Beta24Memory in Genie One — Genie Agents, Beta25Web search in Genie — Genie Agents, Beta26Anomaly detection — Data Quality, Public Preview27Automatic type widening in Auto Loader — Data Ingestion, Public Preview28Managed Iceberg materialized views — Delta Lake, Public Preview29MySQL and PostgreSQL connectors — Data Ingestion, Public Preview30Real-time mode in Lakeflow pipelines — Streaming, Public Preview31Zerobus Ingest into default storage — Data Ingestion, Public Preview32Lakeflow Jobs health score — Runs, Private Preview33AI Runtime — Experiments, Public Preview34AI Runtime and serverless GPU compute — Experiments, Public Preview35Databricks-managed MCP servers — Agents, Public Preview36Feature views — Features, Public Preview37MLflow deployment jobs — Models, Public Preview38MLflow deployment jobs — Models, Public Preview39Agent and MCP services as Unity Catalog securables — Unity Gateway, Beta40Agent services in Unity Catalog — Unity Gateway, Beta41Full-text search indexes in AI Search — AI Search, Beta42Managed agent memory — Agents, Beta43MLflow Prompt Registry — Experiments, Beta44MLflow review queues — Experiments, Beta45Production monitoring for GenAI apps — Experiments, Beta46Production monitoring for GenAI apps — Experiments, Beta47Prompt registry — Experiments, Beta48Service policies for AI securables — Unity Gateway, Beta49Unified trace table — Unity Gateway, Beta50Unity Catalog Skills — Agents, Beta51
written up here on the radar only the beam names each dot as it passes. Hover one to hold its name.
All 51 by number

Written up here the filled blips above

Agent and MCP services as Unity Catalog securables BetaRegistering agents and MCP servers as Unity Catalog objects, so every team's agents are discoverable in one place and every tool a server exposes is chosen, policed and counted.Agent memory BetaLong-term memory for an agent as a Unity Catalog securable, with entries scoped and pathed, governed and audited like a table rather than kept in a side database.AI Runtime and serverless GPU compute PreviewServerless GPU compute for training and fine-tuning: a @distributed decorator in notebooks, the air CLI with a workload YAML, and the successor to Foundation Model Fine-tuning.Feature Views PreviewDeclarative features, defined as a source plus an aggregation over a time window, registered in Unity Catalog and materialised by managed pipelines rather than by a table you build yourself.Production monitoring for GenAI apps BetaRegistered scorers run continuously against a sampled fraction of live traces and attach their verdicts to each trace as feedback, so quality drift shows up without a scheduled evaluation.The Genie Ontology PreviewThe context layer shared by Genie One and Genie Code. Half of it is Unity Catalog semantics that a human governs, half is ranked snippets Genie infers from existing assets.Guardrails for generative applications BetaPolicies evaluated on the way in and on the way out of a model call, returning allow, deny or ask, plus the data-side masking that stops a prompt carrying what it should not.Model Context Protocol on Databricks PreviewThree places an agent's MCP servers come from: Databricks-managed servers, external servers registered in Unity Catalog as MCP Services, and custom servers hosted as apps.MLflow deployment jobs PreviewA deployment job binds a registered model to a Lakeflow job whose evaluation, approval and deployment tasks fire whenever a new model version appears, with a human gate in the middle.Prompt registry BetaThe MLflow prompt registry stores prompt templates as versioned Unity Catalog objects with aliases for production, so a prompt change can be evaluated and rolled back like a model version.Query tags PreviewKey-value tags attached to a session or a single statement, surfacing in query history and the system tables, which is how warehouse spend gets an owner.Domains and Pages PreviewThe human half of the Genie Ontology. Domains group assets by business purpose, Pages define what a business term actually means, and Genie prefers both over what it infers.

What the labels actually promise

StageWho can use itProduction useSupportOn by default
PrivateInvite onlyNoEngineering onlyNo
BetaMost customersNoEngineering onlyPremium yes, Enterprise no
PreviewEveryoneYesThe support teamYes, with an opt-out
GAEveryoneYesThe support teamYes

The line that matters is between Beta and Public Preview. Public Preview is supported and may be used in production. Beta is not, whatever it looks like in the UI. Databricks publishes these definitions on itsrelease types page.

Every entry, with the detail

filter it, or read the radar above
ABAC DENY policies BetaAttribute-based policies that deny rather than grant, evaluated against governed tags.Why it matters: Deny rules change how you design a permission model, so worth watching closely.Sep 2026
Dashboard relationships Previewwe cover itA join graph across AI/BI dashboard datasets, resolved at query time, with cross-dataset measures over several fact tables.Why it matters: The dynamic-grain counterpart to a metric view. Support for relationships in Unity Catalog is still in progress.Sep 2026
Relationship-scoped and cross-dataset filters Previewwe cover itA filter that travels across datasets through the relationships defined on the dashboard.Why it matters: The other half of dashboard relationships, and what makes a multi-fact dashboard behave like one model.Sep 2026
Zerobus Ingest into default storage Previewwe cover itPush records into a Unity Catalog table backed by default storage, with no external location and no message bus.Why it matters: Ingesting into streaming tables is already generally available; this removes the last piece of setup.Sep 2026
Environments on classic compute BetaDeclare dependencies as an environment on a classic cluster instead of cluster libraries and init scripts.Why it matters: One dependency model across serverless and classic compute, instead of two that behave differently.Aug 2026
Fine-grained DML privileges BetaINSERT, UPDATE and DELETE as separate privileges under MODIFY, so a role can change rows without changing schema.Why it matters: The first change to the core Unity Catalog privilege list in years, which dates every permission diagram.Aug 2026
Genie Code task in jobs BetaRuns an autonomous agent from a natural-language prompt as a task inside a job.Why it matters: An agent as an orchestration step raises governance questions worth thinking about before it lands.Aug 2026
Private network gateway PrivateOne managed gateway for serverless traffic to your VPC, to on-premises systems and to the internet, reusing the network connectivity configuration.Why it matters: It would replace today’s sprawl of connectivity configurations, PrivateLink and egress rules with a single object.Aug 2026
REPLACE USING flows BetaKeeps a streaming table in sync from partial snapshots by replacing every row that matches the key columns.Why it matters: Completes the change-capture family next to AUTO CDC and REPLACE WHERE, which are both generally available.Aug 2026
Sharing foreign schemas and tables BetaPut tables that live in another system, federated into Unity Catalog, inside a share.Why it matters: Sharing without copying first. Foreign Delta tables are further ahead, at Public Preview.Aug 2026
The ai_search SQL function BetaRetrieval from an AI Search index, reranking and an optional grounded answer, all from one SQL function.Why it matters: Retrieval-augmented generation collapsed into a single SQL call, with citations arriving alongside it.Aug 2026
Unified trace table BetaEvery request and response across the gateway in one Unity Catalog table, in OpenTelemetry format.Why it matters: The observability half of Unity Gateway reaching general availability.Aug 2026
Unity Catalog Skills BetaGoverned, reusable instructions that an agent can load, registered in Unity Catalog.Why it matters: Skills as a governed object rather than a prompt pasted into a notebook.Aug 2026
Web search in Genie BetaPublic web search with citations inside Genie Code and Genie One, behind one workspace toggle.Why it matters: It puts material from outside the governed catalogue into an answer, which is a governance question as much as a feature.Aug 2026
Git CLI commands in Git folders PreviewRun arbitrary git commands, including stash, interactive rebase and force push, inside a Git folder.Why it matters: Removes the limitation people hit within a week of using Git folders.Jul 2026
Governance Hub BetaAn account-level view of data health, coverage, AI usage and what is driving cost.Why it matters: The new front door for platform administrators, pulling together what is scattered today.Jul 2026
Managed Iceberg materialized views Previewwe cover itMaterialized views stored as managed Iceberg tables, readable by external Iceberg engines.Why it matters: Extends managed Iceberg, generally available since May, from base tables to derived ones.Jul 2026
Memory in Genie One BetaGenie One remembering context between conversations, with confirmation prompts before it keeps something.Why it matters: Changes the experience for business users more than any other addition to the surface.Jul 2026
Unit testing for pipelines BetaPython unit tests over pipeline transforms with mocked table inputs, run from the pipeline editor.Why it matters: The first native testing story for declarative pipelines, and the biggest gap against plain Spark.Jul 2026
Configurable recovery period for dropped tables PreviewSet how long a dropped managed table can be undropped, from nothing at all to thirty days.Why it matters: It changes what UNDROP TABLE can actually do, which is the sort of detail an exam asks about.Jun 2026
Genie Ontology Previewwe cover itThe shared context layer behind Genie, half defined by you in Unity Catalog and half inferred from metric views, dashboards and queries.Why it matters: On by default since August 2026, so it is already shaping answers in workspaces that never opted in.Jun 2026
Lakehouse Replay PreviewShadow-runs a sample of read-only serverless queries against a coming runtime version to catch regressions.Why it matters: It changes how a runtime upgrade is planned, from hoping to measuring.Jun 2026
Real-time mode in Lakeflow pipelines PreviewA trigger setting on a pipeline sink that pushes end-to-end latency down to milliseconds instead of micro-batches.Why it matters: It breaks the micro-batch assumption that sits under everything written about streaming so far.May 2026
Automatic type widening in Auto Loader PreviewEvolves an integer column to long, or float to double, without rewriting the table.Why it matters: Schema evolution is the part of Auto Loader people get wrong most often, and this changes the answer.Mar 2026
Anomaly detection Previewwe cover itSchema-level monitoring that learns freshness and completeness from history and flags tables that went quiet or arrived thin.Why it matters: The only layer that catches an upstream job stopping, which no row-level rule can see.Feb 2026
Agent services in Unity Catalog Betawe cover itAgents registered as Unity Catalog securables, discoverable next to tables and models.Why it matters: Governance for agents that works like governance for data.2026
AI Runtime PreviewServerless GPU compute with a distributed decorator for PyTorch, DeepSpeed and Ray workloads.Why it matters: The replacement for Foundation Model Fine-tuning, which has reached end of life.2026
Bundle deployment history PrivateA record of what each bundle deployment changed, queryable after the fact.Why it matters: Deployment archaeology is currently a git log and a guess.2026
Databricks-managed MCP servers Previewwe cover itReady-made Model Context Protocol servers for Genie, AI Search, SQL and Unity Catalog functions.Why it matters: The documented default way to give an agent tools, ahead of hand-written wrappers.2026
Domains and subdomains Previewwe cover itGroups catalogs, tables and assets into business domains using governed tags, with one level of nesting.Why it matters: The organising layer Unity Catalog has been missing, and it feeds the Genie Ontology.2026
Feature views Previewwe cover itDeclarative features with windowed aggregations, materialised on demand instead of hand-built tables.Why it matters: Databricks now points here first for new feature work.2026
Full-text search indexes in AI Search Betawe cover itIndexes that need no embeddings at all, available on storage-optimized endpoints.Why it matters: The reason Vector Search was renamed AI Search.2026
Genie Inspect Previewwe cover itInspecting how a Genie Agent reached an answer, step by step.Why it matters: The debugging surface for an agent that answers confidently and wrongly.2026
Lakeflow Jobs health score PrivateA health score per job, summarising reliability and cost signals in one number.Why it matters: Worth watching because it would change how job monitoring is presented, not only measured.2026
Lakehouse Real-Time warehouse BetaA fourth SQL warehouse type aimed at low-latency serving, alongside serverless, pro and classic.Why it matters: Changes the answer to "which warehouse type" questions once it settles.2026
Managed agent memory Betawe cover itMemory stores as Unity Catalog securables, holding episodic and semantic memory for an agent.Why it matters: Agents that remember are a different product from agents that do not.2026
MLflow deployment jobs PreviewBind a registered model to a job that evaluates, asks for approval and deploys when a new version appears.Why it matters: Model promotion with a human in the loop, without writing the orchestration yourself.2026
MLflow Prompt Registry Betawe cover itVersion prompt templates like code, with aliases for production and automated optimisation.Why it matters: Prompts are the part of a generative application that changes most and is versioned least.2026
MLflow review queues BetaA reviewer workflow for labelling traces one item at a time, for people who are not workspace users.Why it matters: Expert feedback is the input that makes an LLM judge worth trusting.2026
MySQL and PostgreSQL connectors PreviewManaged log-based change capture from the MySQL binlog and from PostgreSQL logical replication.Why it matters: The two most common open-source operational databases, still behind SQL Server and Oracle in the connector line-up.2026
Pages Betawe cover itGoverned, authoritative definitions of business terms, entities and acronyms, held in Unity Catalog.Why it matters: A glossary that the AI surfaces actually read, rather than a wiki nobody updates.2026
Production monitoring for GenAI apps Betawe cover itRuns the same evaluation scorers against a sample of live traces and attaches the results to each trace.Why it matters: The only honest way to know whether a deployed agent is still good.2026
Query tags Previewwe cover itKey-value tags set per session or per statement that show up in query history and in the system tables.Why it matters: The missing piece for attributing warehouse cost to a team or a dashboard.2026
Serverless compute rate limits PrivateCaps on how much serverless compute an account or workspace can consume.Why it matters: The control that is missing when serverless spend runs away, and there is no cluster to stop.2026
Serverless usage policies PreviewAttribute serverless spend to a team or project by attaching a policy that tags the usage.Why it matters: Serverless usage is otherwise hard to attribute, which is how budgets get surprised.2026
Service policies for AI securables Betawe cover itPolicies evaluated when a model is called and when it answers, returning allow, deny or ask, with built-in safety judges.Why it matters: The replacement for AI guardrails, and it fails closed by design.2026
Agent and MCP services as Unity Catalog securables Betawe cover itRegistering agents and MCP servers as Unity Catalog objects, so every team's agents are discoverable in one place and every tool a server exposes is chosen, policed and counted.
AI Runtime and serverless GPU compute Previewwe cover itServerless GPU compute for training and fine-tuning: a @distributed decorator in notebooks, the air CLI with a workload YAML, and the successor to Foundation Model Fine-tuning.
Production monitoring for GenAI apps Betawe cover itRegistered scorers run continuously against a sampled fraction of live traces and attach their verdicts to each trace as feedback, so quality drift shows up without a scheduled evaluation.
MLflow deployment jobs Previewwe cover itA deployment job binds a registered model to a Lakeflow job whose evaluation, approval and deployment tasks fire whenever a new model version appears, with a human gate in the middle.
Prompt registry Betawe cover itThe MLflow prompt registry stores prompt templates as versioned Unity Catalog objects with aliases for production, so a prompt change can be evaluated and rolled back like a model version.

Labels come from the documentation page of each feature and are re-checked by the agent. A feature that reaches general availability moves out of this page and into the area it belongs to.